priorwork-mcp

MCP server for priorwork.fyi — semantic search over 114,000+ papers from major ML, NLP, and Computer Vision conferences (NeurIPS, CVPR, ICML, ICLR, ACL, EMNLP, ICCV, AISTATS, and more), directly from Claude Code, Cursor, or any MCP client.

Tools

Setup

Claude Code

claude mcp add priorwork -- uvx priorwork-mcp

Cursor / other MCP clients

Add to your MCP configuration:

{
  "mcpServers": {
    "priorwork": {
      "command": "uvx",
      "args": ["priorwork-mcp"]
    }
  }
}

Caching

The topics artifact is ~4.6 MB and changes only when the pipeline runs, so it is cached at ~/.cache/priorwork-mcp/topics.json rather than fetched per search. Set PRIORWORK_CACHE_DIR to move it.

Every search response names the published set it came from, so knowing when to refetch is one string comparison. That covers the change a client cannot infer for itself: a run that rewrites every title leaves every key intact, since a key derives from a topic's supporting papers rather than its text.

A refetch that fails leaves the cached copy in place — stale topics beat no topics, and the id will still not match on the next search, so it retries then. An unwritable cache directory or a corrupt cache file costs a fetch, not an error.

Configuration

Configuration is via environment variables (pass with -e KEY=value in claude mcp add, or an "env" block in JSON configs):

Example

"Find prior work on spatiotemporal forecasting with graph neural networks"

1. Taming Local Effects in Graph-based Spatiotemporal Forecasting — Andrea Cini, et al. (2023 · NeurIPS)  [score: 0.802]
2. FourierGNN: Rethinking Multivariate Time Series Forecasting from a Pure Graph Perspective — ...

Or, for open problems rather than published work:

"We want to apply interpretability methods to models served only behind closed APIs, where activations and full logits are unavailable."

1. A Framework for Deploying Activation- and Logit-Based Methods Against Closed-Source LLM APIs
   Relevance: 0.554   Area: Model Safety, Alignment & Jailbreaks   Evidence: 177 papers
   Why it matters: A very large body of techniques — safety defenses, interpretability
   probes, evaluation metrics — requires hidden states, attention maps, or token-level
   probabilities, and is therefore silently restricted to open-weight models. ...

Describe the problem in a sentence or two. Keywords return nothing: each topic is stored as a synthetic abstract, so efficient transformers edge devices matches nothing while a prose description of the same intent matches well.

License

MIT